- while (mask < num_procs)
- {
- dst = rank ^ mask;
- dst_tree_root = dst >> i;
- dst_tree_root <<= i;
- rank_tree_root = rank >> i;
- rank_tree_root <<= i;
- send_offset = rank_tree_root * send_increment;
- recv_offset = dst_tree_root * recv_increment;
-
- if (dst < num_procs)
- {
- MPI_Sendrecv(tmp_buff + send_offset, curr_size, send_type, dst, tag,
- tmp_buff + recv_offset, mask * recv_count * num_procs,
- recv_type, dst, tag, comm, &status);
-
- MPI_Get_count(&status, recv_type, &last_recv_count);
- curr_size += last_recv_count;
- }
-
-
- if (dst_tree_root + mask > num_procs)
- {
-
- num_procs_completed = num_procs - rank_tree_root - mask;
- /* num_procs_completed is the number of processes in this
- subtree that have all the data. Send data to others
- in a tree fashion. First find root of current tree
- that is being divided into two. k is the number of
- least-significant bits in this process's rank that
- must be zeroed out to find the rank of the root */
-
- j = mask;
- k = 0;
- while (j)
- {
- j >>= 1;
- k++;
- }
- k--;
-
- tmp_mask = mask >> 1;
-
- while (tmp_mask)
- {
- dst = rank ^ tmp_mask;
-
- tree_root = rank >> k;
- tree_root <<= k;
-
- /* send only if this proc has data and destination
- doesn't have data. at any step, multiple processes
- can send if they have the data */
-
- if ((dst > rank)
- && (rank < tree_root + num_procs_completed)
- && (dst >= tree_root + num_procs_completed))
- {
- MPI_Send(tmp_buff + dst_tree_root * send_increment,
- last_recv_count, send_type, dst, tag, comm);
-
- }
-
- /* recv only if this proc. doesn't have data and sender
- has data */
-
- else if ((dst < rank)
- && (dst < tree_root + num_procs_completed)
- && (rank >= tree_root + num_procs_completed))
- {
- MPI_Recv(tmp_buff + dst_tree_root * send_increment,
- mask * num_procs * send_count, send_type, dst,
- tag, comm, &status);
-
- MPI_Get_count(&status, send_type, &last_recv_count);
- curr_size += last_recv_count;
- }
-
- tmp_mask >>= 1;
- k--;
- }
- }
-
- mask <<= 1;
- i++;
+ while (mask < num_procs) {
+ dst = rank ^ mask;
+ dst_tree_root = dst >> i;
+ dst_tree_root <<= i;
+ rank_tree_root = rank >> i;
+ rank_tree_root <<= i;
+ send_offset = rank_tree_root * send_increment;
+ recv_offset = dst_tree_root * recv_increment;
+
+ if (dst < num_procs) {
+ smpi_mpi_sendrecv(tmp_buff + send_offset, curr_size, send_type, dst, tag,
+ tmp_buff + recv_offset, mask * recv_count * num_procs,
+ recv_type, dst, tag, comm, &status);
+
+ last_recv_count = smpi_mpi_get_count(&status, recv_type);
+ curr_size += last_recv_count;
+ }
+
+
+ if (dst_tree_root + mask > num_procs) {
+
+ num_procs_completed = num_procs - rank_tree_root - mask;
+ /* num_procs_completed is the number of processes in this
+ subtree that have all the data. Send data to others
+ in a tree fashion. First find root of current tree
+ that is being divided into two. k is the number of
+ least-significant bits in this process's rank that
+ must be zeroed out to find the rank of the root */
+
+ j = mask;
+ k = 0;
+ while (j) {
+ j >>= 1;
+ k++;
+ }
+ k--;
+
+ tmp_mask = mask >> 1;
+
+ while (tmp_mask) {
+ dst = rank ^ tmp_mask;
+
+ tree_root = rank >> k;
+ tree_root <<= k;
+
+ /* send only if this proc has data and destination
+ doesn't have data. at any step, multiple processes
+ can send if they have the data */
+
+ if ((dst > rank)
+ && (rank < tree_root + num_procs_completed)
+ && (dst >= tree_root + num_procs_completed)) {
+ smpi_mpi_send(tmp_buff + dst_tree_root * send_increment,
+ last_recv_count, send_type, dst, tag, comm);
+
+ }
+
+ /* recv only if this proc. doesn't have data and sender
+ has data */
+
+ else if ((dst < rank)
+ && (dst < tree_root + num_procs_completed)
+ && (rank >= tree_root + num_procs_completed)) {
+ smpi_mpi_recv(tmp_buff + dst_tree_root * send_increment,
+ mask * num_procs * send_count, send_type, dst,
+ tag, comm, &status);
+
+ last_recv_count = smpi_mpi_get_count(&status, send_type);
+ curr_size += last_recv_count;
+ }
+
+ tmp_mask >>= 1;
+ k--;
+ }